DiceDB
Allow Configurable Number of CPU Cores for Multi-Threading in DiceDB
Was going through the code of DiceDB, specifically this snippet:
var numCores int
if config.EnableMultiThreading {
serverErrCh = make(chan error, 1)
numCores = runtime.NumCPU()
logr.Debug("The DiceDB server has started in multi-threaded mode.", slog.Int("number of cores", numCores))
} else {
serverErrCh = make(chan error, 2)
logr.Debug("The DiceDB server has started in single-threaded mode.")
numCores = 1
}
runtime.GOMAXPROCS(numCores)
Currently, DiceDB dynamically configures its threading model based on the EnableMultiThreading flag in the configuration, defaulting to the number of logical CPU cores available via runtime.NumCPU() when multi-threading is enabled. However, this approach can be problematic in containerized environments, such as Kubernetes, where resource constraints (e.g., CPU limits) are enforced per container.
In Kubernetes, a container can be allocated a specific number of CPU cores, and using runtime.NumCPU() may not reflect these limits accurately. For example, if the host machine has 16 cores but a container is only allocated 4 cores, DiceDB would still use all 16 cores unless GOMAXPROCS is manually set. This could lead to resource contention, poor performance, or overuse of allocated resources, resulting in throttling by the container orchestrator.
Issue Summary:
- Current behavior: DiceDB defaults to using all available CPU cores on the host machine (runtime.NumCPU()). Problem: In environments like Kubernetes, containers may have restricted CPU allocations, and the use of runtime.NumCPU() could exceed the allocated resources.
- Proposed solution: Introduce a configuration option allowing the user to specify the number of cores to use dynamically, either through environment variables or a dedicated configuration field. This would ensure DiceDB respects the container's CPU allocation, preventing resource overuse and improving compatibility with cloud-native environments.
I can pick up these changes, can you assign this issue to me @lucifercr07 @JyotinderSingh @AshwinKul28 @arpitbbhayani
@dograprabhav I was thinking that workloads such as databases would run on dedicated nodes. We are designing DiceDB to be running optimally on modern multi-core hardware extracting the max performance out of it. However I agree, we need a more sophisticated logic to run DiceDB that allocating all cores to it blindly. Irrespective this change may be helpful in test environments. @arpitbbhayani @AshwinKul28 @lucifercr07 wdyt?
